[HBASE-16815] Low scan ratio in RPC queue tuning triggers divide by zero exception

• Type: Bug
• Status: Closed
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 1.3.0, 2.0.0
• Fix Version/s: 1.4.0, 1.2.4, 1.1.8, 2.0.0
• Component/s: regionserver, rpc
• Assignee: Guanghao Zhang
• Reporter: Lars George
• Created: 12/Oct/16 11:27
• Updated: 19/Nov/17 07:14
• Resolved: 22/Oct/16 00:09

Description

Trying the following settings:

<property>
  <name>hbase.ipc.server.callqueue.handler.factor</name>
  <value>0.5</value>
</property>
<property>
  <name>hbase.ipc.server.callqueue.read.ratio</name>
  <value>0.5</value>
</property>
<property>
  <name>hbase.ipc.server.callqueue.scan.ratio</name>
  <value>0.1</value>
</property>

With 30 default handlers, this means 15 queues. Further, it means 8 write queues and 7 read queues. 10% of that is 0.7 which is then floor'ed to 0. The debug log confirms it, as the tertiary check omits the scan details when they are zero:

2016-10-12 12:50:27,305 INFO  [main] ipc.SimpleRpcScheduler: Using fifo as user call queue, count=15
2016-10-12 12:50:27,311 DEBUG [main] ipc.RWQueueRpcExecutor: FifoRWQ.default writeQueues=7 writeHandlers=15 readQueues=8 readHandlers=14

But the code in RWQueueRpcExecutor calls RpcExecutor.startHandler() nevertheless and that does this:

    for (int i = 0; i < numHandlers; i++) {
      final int index = qindex + (i % qsize);
      String name = "RpcServer." + threadPrefix + ".handler=" + handlers.size() + ",queue=" +
          index + ",port=" + port;

The modulo triggers then

2016-10-12 11:41:22,810 ERROR [main] master.HMasterCommandLine: Master exiting
java.lang.RuntimeException: Failed construction of Master: class org.apache.hadoop.hbase.master.HMasterCommandLine$LocalHMaster
        at org.apache.hadoop.hbase.util.JVMClusterUtil.createMasterThread(JVMClusterUtil.java:145)
        at org.apache.hadoop.hbase.LocalHBaseCluster.addMaster(LocalHBaseCluster.java:220)
        at org.apache.hadoop.hbase.LocalHBaseCluster.(LocalHBaseCluster.java:155)
        at org.apache.hadoop.hbase.master.HMasterCommandLine.startMaster(HMasterCommandLine.java:222)
        at org.apache.hadoop.hbase.master.HMasterCommandLine.run(HMasterCommandLine.java:137)
        at org.apache.hadoop.util.ToolRunner.run(ToolRunner.java:70)
        at org.apache.hadoop.hbase.util.ServerCommandLine.doMain(ServerCommandLine.java:126)
        at org.apache.hadoop.hbase.master.HMaster.main(HMaster.java:2524)
Caused by: java.lang.ArithmeticException: / by zero
        at org.apache.hadoop.hbase.ipc.RpcExecutor.startHandlers(RpcExecutor.java:125)
        at org.apache.hadoop.hbase.ipc.RWQueueRpcExecutor.startHandlers(RWQueueRpcExecutor.java:178)
        at org.apache.hadoop.hbase.ipc.RpcExecutor.start(RpcExecutor.java:78)
        at org.apache.hadoop.hbase.ipc.SimpleRpcScheduler.start(SimpleRpcScheduler.java:272)
        at org.apache.hadoop.hbase.ipc.RpcServer.start(RpcServer.java:2212)
        at org.apache.hadoop.hbase.regionserver.RSRpcServices.start(RSRpcServices.java:1143)
        at org.apache.hadoop.hbase.regionserver.HRegionServer.(HRegionServer.java:615)
        at org.apache.hadoop.hbase.master.HMaster.(HMaster.java:396)
        at org.apache.hadoop.hbase.master.HMasterCommandLine$LocalHMaster.(HMasterCommandLine.java:312)
        at sun.reflect.NativeConstructorAccessorImpl.newInstance0(Native Method)
        at sun.reflect.NativeConstructorAccessorImpl.newInstance(NativeConstructorAccessorImpl.java:62)
        at sun.reflect.DelegatingConstructorAccessorImpl.newInstance(DelegatingConstructorAccessorImpl.java:45)
        at java.lang.reflect.Constructor.newInstance(Constructor.java:422)
        at org.apache.hadoop.hbase.util.JVMClusterUtil.createMasterThread(JVMClusterUtil.java:140)
        ... 7 more

That causes the server to not even start. I would suggest we either skip the startHandler() call altogether, or make it zero aware.

Another possible option is to reserve at least one scan handler/queue when the scan ratio is greater than zero, but only of there is more than one read handler/queue to begin with. Otherwise the scan handler/queue should be zero and share the one read handler/queue.

Makes sense?

Issue Links

• relates to: HBASE-18937 Backport HBASE-16815 to branch-1.3